In Model Playground, An Agentforce working with a Salesforce-enabled foundational model has control over specific hyperparameters that can directly affect the behavior of the generative model:
* Temperature: Controls the randomness of predictions. A higher temperature leads to more diverse outputs, while a lower temperature makes the model's responses more focused and deterministic.
* Frequency Penalty: Reduces the likelihood of the model repeating the same phrases or outputs frequently.
* Presence Penalty: Encourages the model to introduce new topics in its responses, rather than sticking with familiar, previously mentioned content.
These hyperparameters are adjustable to fine-tune the model's responses, ensuring that it meets the desired behavior and use case requirements. Salesforce documentation confirms that these three are the key tunable hyperparameters in the Model Playground.
For more details, refer to Salesforce AI Model Playground guidance from Salesforce's official documentation on foundational model adjustments.